In this paper, we study quantile distributional policy evaluation, where the return distribution under a given policy is approximated by the quantile Bellman fixed point $\bmη_m$. We investigate the statistical accuracy of estimating $\bmη_m$ across quantile resolutions under Wasserstein distances. Assuming access to a generative model, we construct a quantile fixed point estimator $\bmη_m^{(n)}$ from an empirical Markov decision process of $n$ observations per state-action pair and establish non-asymptotic error bounds under Wasserstein distances. When reward densities are uniformly bounded away from zero, we obtain a root-$n$ high-probability bound simultaneously over all resolutions $m$ and all Wasserstein orders $1\le p\le\infty$. When reward densities vanish at the support endpoints, we show that continuity and endpoint monotonicity suffice to preserve the same resolution-independent rate for $1\le p\le2$. For $p>2$, we establish a root-$n$ bound with coefficient $m^{1/2-1/p}$, together with a slower, resolution-independent $n^{-1/p}$ bound for every finite $p>2$. We further obtain resolution-independent $W_\infty$ consistency governed by the continuity modulus of reward quantile functions. To investigate the sharpness of these bounds, we characterize the exact local asymptotic minimax risk at each fixed resolution as the expected Wasserstein-coordinate norm of the efficient Gaussian limit and prove that our estimator attains it. We then examine the resolution dependence of this optimal local risk. A regular one-state instance has resolution-independent local risk constants, whereas a single endpoint-vanishing two-state instance exhibits growth of order $m^{1/2-1/p}$, up to logarithmic factors along a resolution subsequence, for every $p>2$.
Shuiwen Shen, Jinhua Zhang, Chenlei Hou, Peng Shi, Yong Han, Bo Gao, Xiaoxiao Na, Zhihui Lin 等 9 位作者
Online Active Learning (OAL) aims to extend deep learning from static inference after offline training to continuous adaptation during real-time interaction with a changing physical environment. This transition is essential for safety-critical applications, but it raises a central theoretical question: whether the online parameter update process remains stable when scaled to multilayer neural architectures. This paper develops a stability-focused OAL framework. The framework uses prediction error as a feedback signal to update a structured deep model, thereby linking online learning with Lyapunov-based energy minimization. We show that prediction error convergence alone is insufficient for global parameter convergence because multilayer networks may admit high-dimensional manifolds of equivalent output solutions. The analysis therefore separates output error regulation, internal zero dynamics, interlayer coupling, and parameter convergence. Persistency of excitation and active damping are then used to remove the residual null-space modes that remain after output error convergence. A cascaded Lyapunov analysis establishes asymptotic stability of the overall OAL learning process, while monotonic activation functions and diffeomorphic layer mappings provide the uniqueness condition for the target parameter trajectory. The resulting proof provides a theoretical basis for applying online adaptive learning to large-scale neural systems in nonstationary and safety-critical settings.
Zeroth-order (ZO) optimization provides a gradient-free alternative to first-order (FO) methods by estimating gradients via finite differences of function evaluations, and has recently emerged as a memory-efficient paradigm for fine-tuning large-scale models by avoiding backpropagation. However, ZO optimization has a fundamental tension between accuracy and query efficiency. In this work, we show that ZO optimization can be substantially improved by unifying two complementary principles: (i) a projection-based subspace view that reduces gradient estimation variance by exploiting the intrinsic low-rank structure of model updates, and (ii) Muon-style spectral optimization that applies gradient orthogonalization to extract informative spectral structure from noisy ZO gradients. These findings form a unified framework of subspace gradient orthogonalization, which we instantiate in a new method, ZO-Muon, admitting a natural interpretation as a low-rank Muon optimizer in the ZO setting. Extensive experiments on large language models (LLMs) and vision transformers (ViTs) demonstrate that ZO-Muon significantly accelerates convergence and achieves a win-win improvement in accuracy and query/runtime efficiency. Notably, compared to the popular MeZO baseline, ZO-Muon requires only 24.7% of the queries to reach the same SST-2 performance for LLM fine-tuning, and improves accuracy by 25.1% on ViT-L fine-tuning on CIFAR-100.
Zihan Gu, Lulu Zhao, Tianyi Ma, Hua Zhang, Wenbo Su, Yanan Cao, Yue Hu
Grokking is a delayed transition from memorisation to generalisation. A network first fits observed examples while remaining near chance on held-out inputs, then generalises abruptly. Prior work has focused mainly on this delay, finding that the network's main loss becomes weight decay after the fitting phase. We name the solution to this optimization problem as the minimum-norm endpoint, and study which tasks make all minimum-norm endpoints generalise from partial observations. Specifically, we model pairwise input tasks as edge colourings of $K_{q,q}$, with Latin colourings as our structured family, and classify each observed instance as all-good, mixed, or no-good by its endpoint set. Furthermore, we prove two contrasting results: (1) Finite-abelian-group families are asymptotically all-good from partial observations; (2) Under a uniform random Latin square, all-good tasks become asymptotically rare for observation-based learners, leaving mixed or no-good behaviour. Moreover, experiments show that relabellings of addition generalise despite their dense polynomial representations. Finally, on tasks of composite order, finite-training results suggest that predictors learn a coarse part of the output before the full label.